First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Improvement on 8 Kb/s CELP Using Learned Codebook: LCELP

Toshiki Miyano, Kazunori Ozawa

C&C Information Technology Research Labs., NEC Corporation, Kawasaki, Kanagawa, Japan

This paper proposes 8kb/s LCELP (Learned Code Excited LPC Coding). In order to improve conventional CELP speech quality with relatively low computational complexity, LCELP uses a two-stage vector quantizer with learned and random code-books. The advantages of using both the learned codebook and the random codebook are that synthetic quality is improved by the learned codebook and that robustness for any speech is enhanced by the random codebook. The learned codebook is designed using a speech database, and the random codebook is designed using white Gaussian signals. In order to allocate more bits to the learned and the random codebooks, LSP parameters are efficiently quantized by using a vector-scalar quantization (VQ-SQ) technique. Computer simulation results show that 8 kb/s LCELP achieves an average of 14.5 dB segmental SNR, which is 0.9 dB higher than that for the conventional CELP [2]. Informal listening tests show that LCELP speech quality is high and close to 56 kb/s ...-law PCM.

Full Paper

Bibliographic reference.  Miyano, Toshiki / Ozawa, Kazunori (1990): "Improvement on 8 kb/s CELP using learned codebook: LCELP", In ICSLP-1990, 97-100.